An Efficient CNN-Based Deep Learning Model with Ada-Belief Optimizer for Heart Disease Prediction

Abstract

Heart arrest remains a leading cause for loss of life worldwide, demanding accurate and early prediction to enable timely mediate and minimize healthcare burdens. However, clinical environments often suffer from data noise, missing values, and variability across sources, which significantly degrade the potential of traditional machine learning models. To tackle these constraints, this paper proposes a Deep Ensemble Learning framework for robust heart failure prediction in noisy clinical settings. The proposed model integrates multiple deep learning architectures something like CNNs, LSTMs, and Transformer-based models through a weighted ensemble mechanism to enhance generalization and resilience to noise. Using Ada-Belief optimizer, the system obtained an accuracy of 98 percent after training. Clinical data from publicly available datasets are augmented with synthetic noise to test the model’s robustness. Comparative results indicate that the deep ensemble method significantly overtakes individual models and conventional ensemble practices in terms of accuracy (98.45), sensitivity (99.41), and F1-score (99.81). Furthermore, explainability tools like SHAP are employed to interpret model predictions, ensuring clinical relevance and trust. The findings suggest that deep ensemble learning is a promising avenue for reliable and interpretable heart failure prediction in real-world, noisy healthcare environments.